Who Is the Inventor When AI Helped? Patent Rules After the USPTO's 2025 Reset

The Short Answer

An AI system cannot be named as an inventor on a US patent. An invention made with the help of AI can still be patented, as long as a human conceived what is claimed. In the eyes of the USPTO, AI is a tool, like a lab instrument or a search engine.

That sounds simple, but the guidance behind it changed in late 2025, and plenty of published advice still describes the old test.

What Thaler Settled, and What It Did Not

In Thaler v. Vidal, the Federal Circuit held that the Patent Act requires inventors to be natural persons. It also said it was not deciding whether inventions made by humans with the assistance of AI are patentable (Thaler v. Vidal). The Supreme Court denied review on April 24, 2023 (Docket 22-919).

So Thaler does not mean that using AI defeats a patent. We found no Federal Circuit decision applying conception law to an AI-assisted invention as of this writing, so the USPTO's position is the main guide for now.

The Current Test: Conception by a Human

The USPTO published guidance in February 2024 built around whether a person made a "significant contribution" (89 FR 10043). That approach is gone. Revised guidance published on November 28, 2025 rescinded it entirely and says there is no separate or modified standard for AI-assisted inventions (90 FR 54636).

Under the revised guidance (public inspection copy):

Conception is the touchstone. The question is whether a person possessed knowledge of all the limitations of the claimed invention.
AI systems are instruments. They may provide services and generate ideas, but they remain tools used by human inventors.
The Pannu joint-inventorship factors apply only among humans. They are not used to compare a person to an AI system.
Naming an AI as inventor triggers rejection. That applies to all the claims in the application, and a priority claim to a foreign application naming an AI as sole inventor will not be accepted.

There is an unresolved tension here. The guidance says AI may generate ideas, yet a human must hold every limitation in mind. The Congressional Research Service notes that the guidance is not binding on courts and arguably offers less clarity where AI supplies some or all of the core idea (CRS LSB11251). If an AI produced the key insight and a person only recognized it, ask counsel about it directly.

What You Have to Tell the USPTO

Using AI tools is allowed. The USPTO's April 2024 notice says there is no prohibition on using them and no general duty to tell the office an AI tool helped draft a paper, unless asked. There is also no per se requirement to report AI use in the invention process (89 FR 25609).

The notice does set limits, though parts of it use language from the rescinded 2024 guidance, so read it with the 2025 revision in mind:

A human must sign. Signing cannot be delegated, and AI tools cannot hold USPTO.gov accounts.
Relying on an AI tool's accuracy is not a reasonable inquiry. The signer must review the output, and this applies to inventors filing for themselves too.
The duty of disclosure still applies. Information about AI use must be disclosed if it is material, for example where a drafting tool introduced embodiments the inventors did not conceive and the applicant seeks to patent them.
Reassess inventorship when AI adds to the claims or specification.

The last point is the one most likely to catch teams out. If a drafting tool adds an embodiment and you claim it, ask who conceived it.

Enforcement has already started. In In re Mitchell, a patent attorney received a public reprimand after using generative AI tools to prepare claim constructions for a district court filing that cited intrinsic record material that did not exist or was inaccurately attributed. A second AI tool had been used to review the first (Mitchell Final Order). Checking AI with AI was not enough. A person has to read the sources.

How Other Countries Compare

Stephen Thaler's DABUS applications have tested this worldwide. Offices largely agree that a machine cannot be an inventor. They differ on whether, and how, a human can be named instead.

Jurisdiction	Position
EPO	A machine is not an inventor (J 8/20). A 2026 Board said designation is possible in principle where a person made a causal contribution, but rejected Thaler's self-contradictory filing (T 0528/25)
United Kingdom	An inventor must be a natural person (Thaler v Comptroller). A later attempt naming Thaler himself also failed ([2025] EWHC 2202 (Ch))
Germany	A person can be named where AI was used, as the EPO Board summarizes the Federal Court of Justice ruling (T 0528/25)
Japan	The IP High Court held the right arises only when the inventor is a natural person (IP High Court)
South Africa	The outlier: a grant issued after formal examination only, according to Spoor & Fisher

Thaler's later European and UK losses turned on his own contradictory statements, which suggests that what you say about AI's role can matter as much as the role itself. Run the analysis separately for each country you file in. This area is moving, so check for updates.

Is Pasting an Invention Into an AI Tool a Public Disclosure?

We found no court or patent office decision on this, so anyone who tells you it definitely is, or definitely is not, is getting ahead of the law. Under existing rules, the risk depends on whether the input became available to the public and whether confidentiality applied. US law asks whether the invention was otherwise available to the public (35 USC 102), and the EPO asks whether the public could learn of it with no duty of confidentiality (EPO Guidelines G-IV 1).

Confidentiality deserves attention even without a novelty ruling. The USPTO warns that AI systems may retain information and that providers may suffer breaches. It also warns that servers outside the United States can raise export control issues (89 FR 25609). Separately, a 2026 federal court ruling held chats with a consumer AI tool were not privileged, partly because of the tool's consumer terms, according to Orrick. That is a privilege decision, not a novelty decision, but it shows why a tool's terms matter.

What We Would Document

No US authority prescribes a form for this. These suggestions are ours, based on what the guidance tests:

Who conceived each claim limitation, in a dated note, including what that person understood before and after any AI output
Key prompts and outputs, plus the human selection and testing that followed
Each person's contribution where there are joint inventors
An inventorship recheck whenever claims change
The tool, version, account tier, terms and retention settings you used
An export check before technical data enters a tool
A record that a person read each cited source

Logs can be discoverable and may themselves be material, so have counsel set the retention policy.

The Bottom Line

Use AI, but keep a human at the center of conception and be able to show it. List only people as inventors, read everything before you sign, and choose tools with terms you can defend.

Obsidyan's invention capture and version history can help you keep a structured record of how an invention developed. Who qualifies as an inventor remains a legal question for a registered patent practitioner.

Sources

Sources checked November 6, 2026. Fees, pilot programs and pending rules change, so confirm current figures with the USPTO. This article is general information, not legal advice, and Obsidyan is not a law firm. Talk to a registered patent attorney or agent about your situation.